🤬 The Best, Worst Model
"Opus 5 is one of the best state-of-the-art ultra-smart models and one of the worst state-of-the-art models ever released — because it talks like a complete smartass." The symptoms engineers have been noticing: insanely verbose responses, overuse of phrases like load-bearing, worth stating plainly, here's the honest truth, the em-dash chains — and it burns far more output tokens than any model before it. Plus the Anthropic "co-author" in your Git commit messages. 0:00
⚖️ System Prompt vs User Prompt
There are two ways to prompt your agents. Most engineers use only one. "The user prompt is the single task at hand. The system prompt is the law for every task you hand your agent." 1:58
His method: side-by-side Claude Code instances in a terminal multiplexer (Herder) — "smartass Opus 5" on the left (default system prompt) vs "senior Opus 5" on the right (with an appended system prompt), summarizing Zuckerberg's long "The future is for everyone" post as the demo task. He iterates the system prompt layer by layer. 2:48
He talks to the agent as an engineer, not by assigning a role: "You and I maintain a no-BS, clear, concise, actionable relationship. We're here to solve problems and create value, and our communication reflects that." 5:45
➕➖ Positive + Negative Patterns
The first concrete, repeatable technique: encode both what to replicate and what to avoid. "Replicate the positive patterns, avoid the negative patterns. We're doing both — do this, and don't do this." 8:11
| Positive patterns (do) | Negative patterns (avoid) |
|---|---|
| Put the most important info last — "I always see the last thing you write first" | load-bearing, worth stating plainly, here's the honest truth |
| Use plain, specific language; state each fact once | the real tension, carry the argument |
| Match detail level to the task/request | Analogies — "discuss what's right in front of us" |
| Challenge incorrect assumptions directly and explain why | Overused em-dashes / dash chaining |
| Optimize for clarity and engineering value, not quotability | Flattery, praise, validation, agreement without reason |
| Use the simplest domain terminology that compresses information | Decorative headings, emoji, motivational language, semicolon fragments |
The result after just this section: "very few dashes, no load-bearing anywhere, language clearer, 31 seconds (fewer output tokens)." 12:05
📍 Reference Points
The second technique: build a fast shared language with your agent so it stops repeating tokens. "We use reference points to communicate quickly." 13:20
The payoff: "our agent is referencing R6 — and I can just say 'talk more about R6' and it knows exactly what I mean. We've created this quick language together. It's not repeating itself, not wasting tokens." 15:05
🧱 Hard Operational Boundaries
The third technique targets Opus 5's scope creep — a product of its RL training: "these models are taught to find the answer at all costs, no matter what it takes. Opus 5 will find and reference problems you didn't even ask for remotely. It tries to do as much as possible — and loses focus quickly." 16:50
⌨️ Prompt Aliases
The fourth technique: "micro-skills baked into your system prompt that expand into full instructions on demand." When the agent sees the exact alias, it expands and acts as if the full expansion were given directly. "They're bash aliases, expansions, commands inside your system prompt." 18:38
| Alias | Expansion |
|---|---|
SCR | Simplify, compress, and repeat your response |
ELI | Explain this like I'm 18 — simplify language, shorten response |
FOC | Focus on what matters most — boil it down to the true signal |
REF | Rewrite your response with reference points |
SCR alone triggers it, but a sentence containing "SCR" doesn't. 19:10
The demo: fire SCR at the verbose summary and it compresses instantly; FOC boils it to the single most important point ("signal alignment is being redefined from the model holds the lab's value to the agent holds yours"). 20:50
📋 Examples — In-Context Distillation
The final layer: concrete do/don't examples, "just like training data." The format is user: prompt → do: ideal response → don't: the smartass response. 26:25
The final result: the "senior Opus" response came in at 22 seconds vs 41 seconds for the smartass version, with reference points, no load-bearing, fewer dashes. "If you need to spend output tokens, spend them — but for a concise summary, we don't want all that." 30:35
✅ Key Takeaways
- The system prompt is the law — and the leverage. Every word multiplies across every prompt in and response out. If something should apply globally, that's the place for it.
- Prompt engineering is not dead — it's the highest-leverage skill. "If someone says it's dead, ignore them." Great engineering is great communication, and you communicate with agents through prompts.
- Five repeatable techniques: positive + negative patterns, reference points (D/R/F/P codes), hard operational boundaries, prompt aliases (SCR/ELI/FOC/REF), and in-context-distillation examples.
- Keep the smart, drop the ass. Don't suppress the model's intelligence — just strip the verbal tics, the scope creep, and the token-burning verbosity.
- Talk to it like an engineer, not a role-played persona. "You and I maintain a no-BS, clear, concise, actionable relationship." Models respond better to direct standards than to a fake role.
- You are the bottleneck — not the model. "The hard part now is communicating quickly, concisely, and in the most value-accretive way."
- Write the system prompt by hand, and explain the why. High-multiplication artifacts deserve slow, deliberate work — not vibe-prompting.
🔗 Resources & Links
- 📺 Original video — IndyDevDan's full build
- 💻 Fix Opus 5 codebase — the system prompt + compare harness